Cross-variate conditioning and decoder-depth sensitivity in AdaCast

Investigate cross-variate conditioning for AdaCast's input-conditioned low-rank parameter generation and characterize why AdaCast is more sensitive to updates in early decoder blocks.

Background

AdaCast treats each variable as an independent univariate time series and generates input-specific low-rank updates without modeling dependencies among variables. The authors explicitly identify cross-variate conditioning as unexplored, leaving unresolved whether incorporating relationships among variables can improve the method. They also report that removing early decoder-block updates causes the greatest degradation, but the mechanism behind this sensitivity is not established and requires further study.